See what the
check result means.
By Alvin
Our November 15 walkthrough used Ollama 0.32.0 on the primary Mac, installed model files and the unchanged benchmark. Each brief produced one fresh function checked on four invented-note examples.
| MODEL / BRIEF | EXAMPLES PASSED |
|---|---|
| Qwen3.5 4B / notebook | 4 / 4 |
| Qwen3.5 4B / triage | 4 / 4 |
| Llama3.2 3B / notebook | 1 / 4 |
| Llama3.2 3B / triage | 1 / 4 |
Read one failed example as a reader problem
The input included “Ramp access awaits confirmation.” The notebook should keep it beside the confirmed note; Llama omitted it. For triage, Llama returned strings instead of records with source_id (the note’s label) and quote (its exact words). An editor loses the link back to the source.
The one passing example in each Llama set was the empty list. A 1 / 4 count therefore describes those inputs; it is not a percentage rating of the model’s general ability. Qwen’s 4 / 4 also covers these four examples only.
Practice recovery before a bigger experiment
We stopped the service, read the startup guidance, restarted it and recovered its model list. A deliberately missing model name and missing results file each produced a next action. We installed or removed no weights.
The November 14 tables retain their dates and numbers. This rerun checks the command route and these saved functions. Find its inputs, code and results in lab/beginner-walkthrough-2026-11-15 in the pack.